Enhanced RTK positioning method for blind compensation of laser inertial navigation odometer

Abstract

The invention discloses an enhanced RTK (Real Time Kinematic) positioning method for blind compensation of a laser inertial navigation odometer, which belongs to the field of computer vision and comprises the following specific steps: I, calibrating an initial sensor and performing real-time data fusion; iI, predicting a motion state and carrying out real-time monitoring and evaluation on a GNSS signal; iII, switching a positioning mechanism according to the change condition of the GNSS signal; iV, optimizing a switching process and analyzing radar data in real time to establish an environment model; according to the invention, positioning schemes can be rapidly and accurately switched, positioning errors caused by reduction of signal quality are reduced, the problem of inaccurate positioning caused by unstable GNSS signals is reduced, the robustness and stability of the system in a complex environment are improved, the strain capacity of the whole positioning system is improved, and the positioning accuracy is improved. The positioning error caused by improper path selection is reduced, the strategy can be dynamically adjusted according to the environment change, the dependence on the GNSS signal is reduced, and the response speed and the calculation efficiency of the system are improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to an enhanced RTK positioning method for blind spot filling of a laser inertial navigation odometer. Background Art

[0002] With the rapid development of smart devices and unmanned driving technology, high-precision real-time positioning has become a core element for the safe and efficient operation of various intelligent systems. Current positioning technologies, especially GNSS and its differential enhanced RTK technology, can provide centimeter-level positioning accuracy under good satellite signal conditions. However, GNSS signals are easily blocked or interfered with in complex urban environments, forested areas, tunnels and other scenarios, resulting in a sharp drop in positioning accuracy or even failure. Therefore, how to continue to provide high-precision positioning services when the GNSS signal is unstable or even missing has become a technical problem that needs to be solved urgently; therefore, it is particularly important to invent an enhanced RTK positioning method that fills the blind spot of the laser inertial navigation odometer.

[0003] The existing enhanced RTK positioning method for laser inertial navigation odometer blind spot filling cannot quickly and accurately switch positioning schemes. The positioning error caused by the degradation of signal quality is large, and the problem of inaccurate positioning due to unstable GNSS signals is increased. In addition, the existing enhanced RTK positioning method for laser inertial navigation odometer blind spot filling has poor adaptability of the overall positioning system, increases the positioning error caused by improper path selection, and has a high dependence on GNSS signals, which is not conducive to improving the response speed and computing efficiency of the system. Therefore, we propose an enhanced RTK positioning method for laser inertial navigation odometer blind spot filling. Summary of the invention

[0004] The purpose of the present invention is to solve the defects in the prior art and to propose an enhanced RTK positioning method for laser inertial navigation odometer blind spot filling.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The enhanced RTK positioning method of laser inertial navigation odometer blind spot filling has the following specific steps:

[0007] Ⅰ. Calibrate the initial sensor and perform real-time data fusion;

[0008] II. Predict motion status and conduct real-time monitoring and evaluation of GNSS signals;

[0009] III. Switch positioning mechanism according to changes in GNSS signals;

[0010] IV. Optimize the switching process and analyze radar data in real time to establish an environmental model;

[0011] V. Evaluate the error of multi-sensor data in real time and correct the deviation data;

[0012] VI. Check the GNSS signal recovery and perform mechanism switching and positioning again.

[0013] As a further solution of the present invention, the specific steps of real-time data fusion in step I are as follows:

[0014] S1.1: Acquire each set of historical GPS data, IMU data, and radar data collected by each sensor in real time, pre-process the sensor data through filtering, denoising, and time synchronization, and then build a DNN model. The model structure includes an input layer, multiple sets of hidden layers, and an output layer, and construct an input vector X based on the pre-processed sensor data;

[0015] S1.2: The input layer of the DNN model accepts the input vector X, and then the neurons in each hidden layer process the input vector through the activation function to extract features to learn the correlation between different types of data. At the same time, the extraction result of the previous hidden layer is used as the input of the next hidden layer, and it is iterated layer by layer to optimize the position and posture relationship of each sensor;

[0016] S1.3: The model output layer receives the output of the last hidden layer and generates the optimized data fusion result. The mean square error loss function is used to calculate the difference between the predicted value and the actual value output by the DNN model. The weights and bias vectors of each layer of the DNN model are updated layer by layer through the back propagation algorithm.

[0017] S1.4: The DNN model is continuously trained using multiple sets of sensor data until the mean square error loss function converges to a preset threshold. The training is then stopped and the performance of the trained DNN model is evaluated using unused sensor data. The latest sensor data is then input into the verified DNN model to obtain location information after real-time sensor data fusion.

[0018] As a further solution of the present invention, the GPS data in S1.1 is specifically three-dimensional position coordinates; the IMU data is specifically acceleration and angular velocity data; the radar data is specifically point cloud information;

[0019] The specific calculation formula of the root mean square error loss function described in S1.3 is as follows:

[0020]

[0021] Where L represents the difference between the actual value and the predicted value; n represents the total number of sensor data; y i represents the actual fusion result of the i-th group of sensor data; Represents the prediction fusion result of the i-th group of sensor data of the DNN model.

[0022] As a further solution of the present invention, the specific steps of predicting the motion state in step II are as follows:

[0023] S2.1: Construct the corresponding state space S based on the current position, speed and direction data of the object, and represent the state of the object at time t with s t Indicates that, according to the acceleration, deceleration and turning actions taken by the current object in each state, the corresponding action space A is generated, and the action taken by the object at time t is represented by a t express;

[0024] S2.2: Calculate and execute any action a t After that, the object changes from state s t Transfer to another state s t+1 The transfer probability is calculated by detecting potential obstacles and adjusting the vehicle path through radar and IMU data to evaluate the object's action a t The quality of the final state is evaluated by R(s t ,a t )express;

[0025] S2.3: Based on the fusion results of GPS data, IMU data and radar data, the current state s of the object is calculated recursively t Next, perform the optimal action a t The cumulative reward obtained is obtained based on the state transition probability and the cumulative reward result in each state s t The best action a under t , and take it as the optimal strategy;

[0026] S2.4: Collect object motion data collected by various sensors, and predict the object's position information, speed and direction at the next moment based on the optimal action selected under the object's current state, and compare and correct it based on the state prediction and actual observation data.

[0027] As a further solution of the present invention, the specific steps of the optimized switching process in step IV are as follows:

[0028] S3.1: Real-time monitoring of the signal-to-noise ratio (SNR), number of received satellites, and position accuracy (HDOP) of the GNSS signal to obtain the GNSS signal quality (Q = f(SNR, number of satellites, HDOP). When the signal quality (Q) drops to a preset threshold (Q threshold When the following occurs, the positioning mechanism is switched and initial weights are assigned to each group of candidate strategies based on historical data or experience settings;

[0029] S3.2: Based on the initial weight information of each group of candidate strategies, the initial quantum states of different sensor combinations are obtained at the initial state, and the state space of the candidate strategies is constructed according to the different combinations in the corresponding quantum states. Then, according to the interaction and influence between the sensor combinations in the actual scene, the quantum state evolution is performed, and the weights of each sensor are calculated in real time;

[0030] S3.3: During the evolution process, when the radar data is more stable than the GPS data, the probability of selecting the radar solution increases; when the GPS signal recovers and becomes stable, the probability of selecting the GPS solution increases; when the GPS signal is weak or the radar signal deviation value exceeds the preset threshold, the probability of selecting the IMU solution increases;

[0031] S3.4: When the positioning error of the fusion solution is reduced to the preset threshold range, the evolution is stopped, and the optimal sensor combination solution is determined through quantum measurement. At the same time, the obtained optimal sensor combination solution replaces the original solution. When the GNSS signal changes again, it is judged again, and a new sensor combination is searched and optimized.

[0032] As a further solution of the present invention, the candidate strategies described in S3.1 specifically include positioning relying only on GPS, positioning by GPS and radar, tightly coupled positioning by IMU and radar, and fusion positioning of GPS, radar and IMU;

[0033] The specific expression formula of the quantum state described in S3.2 is as follows:

[0034] ψ>=α GPS |GPS>+α radar |radar>+α IMU |IMU>

[0035] Where ψ> represents the superposition state of candidate switching strategies; α GPS |GPS> represents the state of using GPS data in the sensor combination scheme, where α GPS represents the probability of selecting the GPS solution; α radar |radar> represents the state of using radar data in the sensor combination scheme, where α radar represents the probability of selecting the radar solution; α IMU |IMU> represents the state of using IMU data in the sensor combination scheme, where α IMU Represents the probability of selecting the IMU solution.

[0036] As a further solution of the present invention, the specific steps of analyzing radar data to establish an environment model in step IV are as follows:

[0037] S4.1: When switching to tightly coupled positioning through IMU and radar, the radar sensor continuously scans the surrounding environment, collects 3D point cloud data, and constructs a corresponding undirected weighted graph G = (V, E) based on the collected 3D point cloud data, where each point v in the 3D point cloud data l ∈V, represents obstacles in the environment, and the edge e between two points lm ∈E, represents the spatial distance between two points;

[0038] S4.2: Select any point in the undirected weighted graph as the starting point, add the starting point to the tree structure, build an initial tree, and set the weight of the starting point. Then, select the shortest edge connecting any point in the current tree structure and a vertex that has not been added to the tree from the undirected weighted graph, and add the corresponding vertex to the tree structure;

[0039] S4.3: After the vertex is added, the weights of other vertices connected to the vertex that have not been added to the tree are updated. Then, a shortest edge connecting the vertices that has not been added to the tree is continuously selected, and the vertex corresponding to the edge is added to the tree. The edge weights between the vertex and other vertices that have not been added to the tree are updated until all vertices are added to the tree, and the establishment of the environment model is completed;

[0040] S4.4: Based on the acceleration and angular velocity provided by the IMU data and the predicted position information at the next moment, the corresponding displacement and attitude change are obtained. Combined with the IMU prediction results and the radar's environmental model, the current position information is updated through Kalman filtering, and the updated position information is fed back to update the object state.

[0041] As a further solution of the present invention, the specific calculation formula of the Kalman filter in S4.4 is as follows:

[0042]

[0043] In the formula, Represents the position information at time k after combining IMU data and radar data; Represents the position information at time k-1 after combining IMU data and radar data; K k represents the Kalman gain at time k and is dynamically adjusted according to the observation error of IMU data and radar data; z k represents the actual observation value of the radar data at time k; Represents the predicted observation value at time k of the IMU data.

[0044] As a further solution of the present invention, the specific steps of correcting the deviation data in step V are as follows:

[0045] S5.1: Initialize GPS data, radar data and IMU data, extract the position information, attitude and motion data of each data, and measure the deviation of different sensors through the error evaluation function, and then generate the corresponding number of correction bodies according to the number of correction strategies;

[0046] S5.2: Construct the corresponding error space according to the calculation result of the error evaluation function, randomly assign the correction strategy as the starting point of the correction body, and initialize the pheromone concentration of each correction strategy. According to the pheromone concentration and the error of the current sensor data, calculate the probability of the correction body moving to the next correction strategy, and randomly select the next new correction strategy to move according to the selection probability;

[0047] S5.3: Obtain the errors of GPS data, radar data and IMU data under the new strategy through the error evaluation function. If the sensor data error is lower than the preset threshold, the pheromone of the selected path is increased, otherwise the pheromone of the path is reduced, and the correction of body position and pheromone update are repeated;

[0048] S5.4: When the error evaluation function value variation range reaches the preset convergence range, stop updating, and select the correction strategy with the smallest error among GPS data, radar data and IMU data as the optimal correction strategy. Based on this correction strategy, correct each group of data and output the fused data with the smallest error.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. The enhanced RTK positioning method for blind spot filling of the laser inertial navigation odometer monitors the signal-to-noise ratio (SNR) of the GNSS signal, the number of received satellites and the position accuracy (HDOP) in real time to obtain the quality of the GNSS signal. When the signal quality drops below the preset threshold, the positioning mechanism is switched, and the initial weight is assigned to each group of candidate strategies according to historical data or experience. According to the initial weight information of each group of candidate strategies, the initial quantum state of different sensor combinations is obtained at the initial state, and the state space of the candidate strategy is constructed according to different combinations in the corresponding quantum state. Then, according to the interaction and influence between the sensor combinations in the actual scene, the quantum state is evolved, and the weight of each sensor is adjusted in real time. During the evolution process, when the radar data is more stable than the GPS data, the radar solution is selected. The probability of selecting the GPS solution increases when the GPS signal returns to stability. When the GPS signal is weak or the radar signal deviation value exceeds the preset threshold, the probability of selecting the IMU solution increases. When the positioning error of the fusion solution is reduced to the preset threshold range, the evolution is stopped, and the optimal sensor combination solution is determined through quantum measurement. At the same time, the obtained optimal sensor combination solution replaces the original solution. When the GNSS signal changes again, it is judged again, and the new sensor combination is searched and optimized. The positioning solution can be switched quickly and accurately, reducing the positioning error caused by the decline in signal quality, reducing the inaccurate positioning problem caused by unstable GNSS signals, improving the robustness and stability of the system in complex environments, and improving the overall positioning accuracy.

[0051] 2. The enhanced RTK positioning method for blind spot filling of the laser inertial navigation odometer constructs a corresponding state space according to the current position, speed and direction data of the object, and then generates a corresponding action space according to the acceleration, deceleration and turning actions taken by the current object in each state, calculates the transfer probability of the object from the current state to another state after executing any action, and evaluates the state of the object after executing the action through potential obstacles detected by radar and IMU data and vehicle adjustment paths. According to the fusion results of GPS data, IMU data and radar data, the cumulative reward obtained when the object performs the optimal action in the current state is calculated in a recursive manner, and the optimal action in each state is obtained according to the state transfer probability and the cumulative reward result, and it is used as the optimal strategy to collect the object motion data collected by various sensors, and based on the optimal action selected in the current state of the object, predict the position information, speed and direction of the object at the next moment, and compare and correct it according to the state prediction and actual observation data, so as to improve the resilience of the overall positioning system, reduce the positioning error caused by improper path selection, dynamically adjust the strategy according to environmental changes, reduce dependence on GNSS signals, and help improve the response speed and computing efficiency of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0053] Figure 1 This is a flowchart of the enhanced RTK positioning method for blind spot filling of the laser inertial navigation odometer proposed in the present invention. DETAILED DESCRIPTION

[0054] Example 1

[0055] Reference Figure 1 , the enhanced RTK positioning method of laser inertial navigation odometer blind spot filling, the specific steps of this positioning method are as follows:

[0056] Initial sensor calibration and real-time data fusion.

[0057] Specifically, each set of historical GPS data, IMU data, and radar data collected by each sensor is acquired in real time, and the sensor data is preprocessed by filtering, denoising, and time synchronization, and then a DNN model is constructed. The model structure includes an input layer, multiple sets of hidden layers, and an output layer. An input vector X is constructed based on the preprocessed sensor data. The DNN model input layer accepts the input vector X, and then the neurons in each hidden layer process the input vector through an activation function to extract features to learn the correlation between different types of data. At the same time, the extraction result of the previous hidden layer is used as the input of the next hidden layer, and it is iterated layer by layer to optimize each sensor. The model output layer receives the output of the last hidden layer and generates an optimized data fusion result. The difference between the predicted value and the actual value output by the DNN model is calculated through the mean square error loss function. The weights and bias vectors of each layer of the DNN model are updated layer by layer through the back propagation algorithm. The DNN model is continuously trained with multiple sets of sensor data until the mean square error loss function converges to the preset threshold. The training is stopped and the performance of the trained DNN model is evaluated with unused sensor data. The latest sensor data is then input into the verified DNN model to obtain the position information after real-time sensor data fusion.

[0058] In this embodiment, the GPS data is specifically three-dimensional position coordinates; the IMU data is specifically acceleration and angular velocity data; the radar data is specifically point cloud information;

[0059] The specific calculation formula of the root mean square error loss function is as follows:

[0060]

[0061] Where L represents the difference between the actual value and the predicted value; n represents the total number of sensor data; y irepresents the actual fusion result of the i-th group of sensor data; Represents the prediction fusion result of the i-th group of sensor data of the DNN model.

[0062] Predict motion status and monitor and evaluate GNSS signals in real time.

[0063] Specifically, the corresponding state space S is constructed according to the current position, speed, and direction data of the object, and the state of the object at time t is represented by s t Indicates that, according to the acceleration, deceleration and turning actions taken by the current object in each state, the corresponding action space A is generated, and the action taken by the object at time t is represented by a t Indicates that the calculation executes any action a t After that, the object changes from state s t Transfer to another state s t+1 The transfer probability is calculated by detecting potential obstacles and adjusting the vehicle path through radar and IMU data to evaluate the object's action a t The quality of the final state is evaluated by R(s t ,a t ) indicates that the current state s of the object is calculated recursively based on the fusion results of GPS data, IMU data and radar data. t Next, perform the optimal action a t The cumulative reward obtained is obtained based on the state transition probability and the cumulative reward result in each state s t The best action a under t , and use it as the optimal strategy, collect object motion data collected by various sensors, and predict the object's position information, speed and direction at the next moment based on the optimal action selected under the object's current state, and compare and correct it based on the state prediction and actual observation data.

[0064] The positioning mechanism is switched according to the changes in GNSS signals.

[0065] Example 2

[0066] Reference Figure 1 , the enhanced RTK positioning method of laser inertial navigation odometer blind spot filling, the specific steps of this positioning method are as follows:

[0067] Optimize the switching process and analyze radar data in real time to build an environmental model.

[0068] Specifically, the signal-to-noise ratio (SNR), number of received satellites and position accuracy (HDOP) of the GNSS signal are monitored in real time to obtain the GNSS signal quality (Q = f(SNR, number of satellites, HDOP). When the signal quality (Q) drops to a preset threshold (Q threshWhen the value is less than old, the positioning mechanism is switched, and initial weights are assigned to each group of candidate strategies according to historical data or experience. According to the initial weight information of each group of candidate strategies, the initial quantum states of different sensor combinations are obtained when the initial state is obtained, and the state space of the candidate strategies is constructed according to the different combinations in the corresponding quantum states. Then, according to the interaction and influence between the sensor combinations in the actual scene, the quantum state evolution is carried out, and the weights of each sensor are adjusted in real time. During the evolution process, when the radar data is more stable than the GPS data, the probability of selecting the radar solution increases. When the GPS signal recovers to stability, the probability of selecting the GPS solution increases. When the GPS signal is weak or the radar signal deviation value exceeds the preset threshold, the probability of selecting the IMU solution increases. When the positioning error of the fusion solution is reduced to the preset threshold range, the evolution is stopped, and the optimal sensor combination solution is determined by quantum measurement. At the same time, the obtained optimal sensor combination solution replaces the original solution. When the GNSS signal changes again, it is judged again, and a new sensor combination is searched and optimized.

[0069] Specifically, when switching to tightly coupled positioning through IMU and radar, the radar sensor continuously scans the surrounding environment, collects 3D point cloud data, and constructs a corresponding undirected weighted graph G = (V, E) based on the collected 3D point cloud data, where each point v in the 3D point cloud data l ∈V, represents obstacles in the environment, and the edge e between two points lm ∈E, represents the spatial distance between two points, selects any point in the undirected weighted graph as the starting point, and adds the starting point to the tree structure, constructs the initial tree, and sets the weight of the starting point. Then, selects the shortest edge connecting any point in the current tree structure and the vertex that has not been added to the tree from the undirected weighted graph, and adds the corresponding vertex to the tree structure. After the vertex is added, the weights of other vertices connected to the vertex that have not been added to the tree are updated. Then, a shortest edge connecting vertices that has not been added to the tree is continuously selected, and the vertex corresponding to the edge is added to the tree, and the edge weights with other vertices that have not been added to the tree are updated until all vertices are added to the tree. The establishment of the environmental model is completed. According to the acceleration and angular velocity provided by the IMU data, and based on the predicted position information at the next moment, the corresponding displacement and attitude change are obtained. Combined with the IMU prediction results and the radar's environmental model, the current position information is updated through the Kalman filter, and the updated position information is fed back to update the object state.

[0070] It should be further explained that the candidate strategies specifically include positioning relying solely on GPS, positioning through GPS and radar, tightly coupled positioning through IMU and radar, and fusion positioning of GPS, radar, and IMU;

[0071] The specific expression formula of quantum state is as follows:

[0072] |ψ>=αGPS |GPS>+α radar |radar>+α IMU |IMU>

[0073] Where |ψ> represents the superposition state of candidate switching strategies; α GPS |GPS> represents the state of using GPS data in the sensor combination scheme, where α GPS represents the probability of selecting the GPS solution; α radar |radar> represents the state of using radar data in the sensor combination scheme, where α radar represents the probability of selecting the radar solution; α IMU |IMU> represents the state of using IMU data in the sensor combination scheme, where α IMU Represents the probability of selecting the IMU solution;

[0074] The specific calculation formula of Kalman filter is as follows:

[0075]

[0076] In the formula, Represents the position information at time k after combining IMU data and radar data; Represents the position information at time k-1 after combining IMU data and radar data; K k represents the Kalman gain at time k and is dynamically adjusted according to the observation error of IMU data and radar data; z k represents the actual observation value of the radar data at time k; Represents the predicted observation value at time k of the IMU data.

[0077] Evaluate the error of multi-sensor data in real time and correct the biased data.

[0078] Specifically, initialize GPS data, radar data and IMU data, extract the position information, posture and motion data of each data, and measure the deviation of different sensors through the error evaluation function. Then, generate a corresponding number of correction bodies according to the number of correction strategies, construct the corresponding error space according to the calculation result of the error evaluation function, randomly assign correction strategies as the starting point of the correction body, and initialize the pheromone concentration of each correction strategy. According to the pheromone concentration and the error of the current sensor data, calculate the probability that the correction body moves to the next correction strategy, and according to the selection probability, randomly select the next new correction strategy to move, and obtain the error of GPS data, radar data and IMU data under the new strategy through the error evaluation function. If the sensor data error is lower than the preset threshold, increase the pheromone of the selected path, otherwise reduce the pheromone of the path, repeat the correction body position and pheromone update, and stop updating when the error evaluation function value variation range reaches the preset convergence range, and select the correction strategy with the smallest error of GPS data, radar data and IMU data as the optimal correction strategy, and calibrate each group of data based on the correction strategy, and output the fusion data with the smallest error.

[0079] Detect GNSS signal recovery and perform mechanism switching and positioning again.

[0080] In this embodiment, the preprocessed GPS data and radar data are converted to the same coordinate system, and the corresponding points of the objects measured by the radar in the GPS coordinate system are found. The coordinate system conversion relationship and time synchronization are optimized through multiple iterations to gradually reduce the error between the radar and GPS data. After the calibration is completed, the calibration accuracy of the radar and GPS is verified using an independent data set, and the position and posture of the object are reflected in real time based on the verified calibration data.

Claims

1. The enhanced RTK positioning method for laser inertial navigation odometer blind spot filling is characterized by: The specific steps of this positioning method are as follows: Ⅰ. Calibrate the initial sensor and perform real-time data fusion; II. Predict motion status and conduct real-time monitoring and evaluation of GNSS signals; III. Switch positioning mechanism according to changes in GNSS signals; IV. Optimize the switching process and analyze radar data in real time to establish an environmental model; V. Evaluate the error of multi-sensor data in real time and correct the deviation data; VI. Check the GNSS signal recovery and perform mechanism switching and positioning again.

2. The enhanced RTK positioning method for laser inertial navigation odometer blind spot filling according to claim 1 is characterized in that: The specific steps of real-time data fusion described in step Ⅰ are as follows: S1.1: Acquire each set of historical GPS data, IMU data, and radar data collected by each sensor in real time, pre-process the sensor data through filtering, denoising, and time synchronization, and then build a DNN model. The model structure includes an input layer, multiple sets of hidden layers, and an output layer, and construct an input vector X based on the pre-processed sensor data; S1.2: The input layer of the DNN model accepts the input vector X, and then the neurons in each hidden layer process the input vector through the activation function to extract features to learn the correlation between different types of data. At the same time, the extraction result of the previous hidden layer is used as the input of the next hidden layer, and it is iterated layer by layer to optimize the position and posture relationship of each sensor; S1.3: The model output layer receives the output of the last hidden layer and generates the optimized data fusion result. The mean square error loss function is used to calculate the difference between the predicted value and the actual value output by the DNN model. The weights and bias vectors of each layer of the DNN model are updated layer by layer through the back propagation algorithm. S1.4: The DNN model is continuously trained using multiple sets of sensor data until the mean square error loss function converges to a preset threshold. The training is then stopped and the performance of the trained DNN model is evaluated using unused sensor data. The latest sensor data is then input into the verified DNN model to obtain location information after real-time sensor data fusion.

3. The enhanced RTK positioning method for laser inertial navigation odometer blind spot filling according to claim 2 is characterized in that: The GPS data in S1.1 is specifically the three-dimensional position coordinates; the IMU data is specifically the acceleration and angular velocity data; the radar data is specifically the point cloud information; The specific calculation formula of the root mean square error loss function described in S1.3 is as follows: Where L represents the difference between the actual value and the predicted value; n represents the total number of sensor data; y i represents the actual fusion result of the i-th group of sensor data; Represents the prediction fusion result of the i-th group of sensor data of the DNN model.

4. The enhanced RTK positioning method for laser inertial navigation odometer blind spot filling according to claim 2 is characterized in that: The specific steps of predicting the motion state described in step II are as follows: S2.1: Construct the corresponding state space S based on the current position, speed and direction data of the object, and represent the state of the object at time t with s t Indicates that, according to the acceleration, deceleration and turning actions taken by the current object in each state, the corresponding action space A is generated, and the action taken by the object at time t is represented by a t express; S2.2: Calculate and execute any action a t After that, the object changes from state s t Transfer to another state s t+1 The transfer probability is calculated by detecting potential obstacles and adjusting the vehicle path through radar and IMU data to evaluate the object's action a t The quality of the final state is evaluated by R(s t ,a t )express; S2.3: Based on the fusion results of GPS data, IMU data and radar data, the current state s of the object is calculated recursively t Next, perform the optimal action a t The cumulative reward obtained is obtained based on the state transition probability and the cumulative reward result in each state s t The best action a under t , and take it as the optimal strategy; S2.4: Collect object motion data collected by various sensors, and predict the object's position information, speed and direction at the next moment based on the optimal action selected under the object's current state, and compare and correct it based on the state prediction and actual observation data.

5. The enhanced RTK positioning method for laser inertial navigation odometer blind spot filling according to claim 4 is characterized in that: The specific steps of the optimized switching process described in step IV are as follows: S3.1: Real-time monitoring of the signal-to-noise ratio (SNR), number of received satellites, and position accuracy (HDOP) of the GNSS signal to obtain the GNSS signal quality (Q = f(SNR, number of satellites, HDOP). When the signal quality (Q) drops to a preset threshold (Q threshold When the following occurs, the positioning mechanism is switched and initial weights are assigned to each group of candidate strategies based on historical data or experience settings; S3.2: Based on the initial weight information of each group of candidate strategies, the initial quantum states of different sensor combinations are obtained at the initial state, and the state space of the candidate strategies is constructed according to the different combinations in the corresponding quantum states. Then, according to the interaction and influence between the sensor combinations in the actual scene, the quantum state evolution is performed, and the weights of each sensor are calculated in real time; S3.3: During the evolution process, when the radar data is more stable than the GPS data, the probability of selecting the radar solution increases; when the GPS signal recovers and becomes stable, the probability of selecting the GPS solution increases; when the GPS signal is weak or the radar signal deviation value exceeds the preset threshold, the probability of selecting the IMU solution increases; S3.4: When the positioning error of the fusion solution is reduced to the preset threshold range, the evolution is stopped, and the optimal sensor combination solution is determined through quantum measurement. At the same time, the obtained optimal sensor combination solution replaces the original solution. When the GNSS signal changes again, the judgment is made again, and the new sensor combination is searched and optimized.

6. The enhanced RTK positioning method for laser inertial navigation odometer blind spot filling according to claim 5 is characterized in that: The candidate strategies described in S3.1 specifically include positioning based on GPS alone, positioning based on GPS and radar, tightly coupled positioning based on IMU and radar, and fusion positioning of GPS, radar, and IMU; The specific expression formula of the quantum state described in S3.2 is as follows: ψ>=a GPS |GPS>+a radar |radar>+a IMU |IMU> Where ψ> represents the superposition state of candidate switching strategies; α GPS |GPS> represents the state of using GPS data in the sensor combination scheme, where α GPS represents the probability of selecting the GPS solution; α radar |radar> represents the state of using radar data in the sensor combination scheme, where α radar represents the probability of selecting the radar solution; α IMU |IMU> represents the state of using IMU data in the sensor combination scheme, where α IMU Represents the probability of selecting the IMU solution.

7. The enhanced RTK positioning method for laser inertial navigation odometer blind spot filling according to claim 5 is characterized in that: The specific steps of analyzing radar data and establishing an environment model in step IV are as follows: S4.1: When switching to tightly coupled positioning through IMU and radar, the radar sensor continuously scans the surrounding environment, collects 3D point cloud data, and constructs a corresponding undirected weighted graph G = (V, E) based on the collected 3D point cloud data, where each point v in the 3D point cloud data l ∈V, represents obstacles in the environment, and the edge e between two points lm ∈E, represents the spatial distance between two points; S4.2: Select any point in the undirected weighted graph as the starting point, add the starting point to the tree structure, build an initial tree, and set the weight of the starting point. Then, select the shortest edge connecting any point in the current tree structure and a vertex that has not been added to the tree from the undirected weighted graph, and add the corresponding vertex to the tree structure; S4.3: After the vertex is added, the weights of other vertices connected to the vertex that have not been added to the tree are updated. Then, a shortest edge connecting the vertices that has not been added to the tree is continuously selected, and the vertex corresponding to the edge is added to the tree. The edge weights between the vertex and other vertices that have not been added to the tree are updated until all vertices are added to the tree, and the establishment of the environment model is completed; S4.4: Based on the acceleration and angular velocity provided by the IMU data and the predicted position information at the next moment, the corresponding displacement and attitude change are obtained. Combined with the IMU prediction results and the radar's environmental model, the current position information is updated through Kalman filtering, and the updated position information is fed back to update the object state.

8. The enhanced RTK positioning method for laser inertial navigation odometer blind spot filling according to claim 7 is characterized in that: The specific calculation formula of Kalman filtering described in S4.4 is as follows: In the formula, Represents the position information at time k after combining IMU data and radar data; Represents the position information at time k-1 after combining IMU data and radar data; K k represents the Kalman gain at time k and is dynamically adjusted according to the observation error of IMU data and radar data; z k represents the actual observation value of the radar data at time k; Represents the predicted observation value at time k of the IMU data.

9. The enhanced RTK positioning method for laser inertial navigation odometer blind spot filling according to claim 1, characterized in that: The specific steps of correcting the deviation data in step V are as follows: S5.1: Initialize GPS data, radar data and IMU data, extract the position information, attitude and motion data of each data, and measure the deviation of different sensors through the error evaluation function, and then generate the corresponding number of correction bodies according to the number of correction strategies; S5.2: Construct the corresponding error space according to the calculation result of the error evaluation function, randomly assign the correction strategy as the starting point of the correction body, and initialize the pheromone concentration of each correction strategy. According to the pheromone concentration and the error of the current sensor data, calculate the probability of the correction body moving to the next correction strategy, and randomly select the next new correction strategy to move according to the selection probability; S5.3: Obtain the errors of GPS data, radar data and IMU data under the new strategy through the error evaluation function. If the sensor data error is lower than the preset threshold, the pheromone of the selected path is increased, otherwise the pheromone of the path is reduced, and the correction of body position and pheromone update are repeated; S5.4: When the error evaluation function value variation range reaches the preset convergence range, stop updating, and select the correction strategy with the smallest error among GPS data, radar data and IMU data as the optimal correction strategy. Based on this correction strategy, correct each group of data and output the fused data with the smallest error.

Images